Score the workflow, not the excitement
List workflows at the level of a repeatable job: classify an inbound request, summarize a call, draft a first response from approved knowledge, reconcile campaign names, or prepare a weekly report. Avoid broad labels like automate marketing or build an agent.
Score each workflow on frequency, time per run, repeatability, data access, quality measurability, acceptable error cost, human review, and whether employees will use the result. High-frequency, lower-risk work with clear patterns is usually the strongest starting point.
- The trigger and desired output are specific
- Representative inputs are available
- A reviewer can recognize good and bad output
- Failures can be contained without customer or business harm
- The team has a real owner for the workflow
Measure the current process before estimating savings
Record monthly volume, active work time, wait time, rework, handoffs, error frequency, escalation, and the value of faster completion. Separate time that can actually be redeployed from theoretical minutes saved.
Include the current quality level. An automation that produces faster drafts but doubles review time or creates subtle errors may move cost rather than remove it.
Model the full operating cost
Add discovery, design, integration, data preparation, model usage, hosting, observability, security review, evaluation, human review, training, maintenance, and expected correction cost. Account for volume growth and for changes in model prices or provider behavior.
The initial build is only one part of the cost. Knowledge changes, APIs break, workflows drift, and teams create exceptions. Budget for ownership after the prototype.
- One-time design and integration cost
- Variable model, tool, and infrastructure cost
- Human review and exception handling
- Evaluation, monitoring, and maintenance
- Expected cost of errors and control failures
- Training and adoption effort
Pilot with a quality gate and a stopping rule
Build a small evaluation set from representative work, including difficult and unsafe cases. Define pass criteria for correctness, completeness, tone, citation, latency, cost, and escalation before evaluating the prototype.
Run the system beside the existing workflow long enough to observe corrections and adoption. Stop or redesign when quality, economics, privacy, or user behavior does not support expansion. A pilot that prevents a poor rollout is still valuable.
Measure durable value after launch
Track usage, completion, review rate, correction severity, time to outcome, throughput, cost per run, failure categories, and the business outcome the workflow supports. Compare the same baseline over a meaningful operating period.
Expansion should follow stable evidence. Add more volume, permissions, data, or autonomy one controlled step at a time, with an owner who can pause the system when the environment changes.